fix: align denoising score target with energy gradient - #333

Problem

BaseScoreMatching.compute_score() returns grad E, but denoising score matching compared it with the conditional log-density score -noise / sigma^2. The opposite conventions make DSM train a quadratic energy in the wrong direction.

Solution

Match the energy gradient against noise / sigma^2, the negative of the conditional log-density score.

The regression test starts a closed-form quadratic energy at zero scale. Standard normal data should push that scale positive, so the loss gradient must be negative. The test applies the same check to exact, denoising, and sliced score matching, pinning their shared convention. On the original code only the DSM case fails (grad = +1.9391); after this change all three pass.

Validation

• pytest tests/losses/test_denoising_score_matching.py -v — 9 passed, 1 CUDA skip
• pytest tests/ -v — 1764 passed, 108 skipped, 21 deselected
• git diff --check

Closes #318